Unsupervised Learning of Object Landmarks via Self-Training Correspondence

Authors: Dimitrios Mallis, Enrique Sanchez, Matthew Bell, Georgios Tzimiropoulos

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental This Section presents experiments illustrating the results produced by our method and by recent state-of-the-art approaches, as well as ablation studies shedding light into some of the key properties of our method. Datasets: We validate our approach on faces, human bodies and cat faces.
Researcher Affiliation Collaboration Dimitrios Mallis University of Nottingham dimitrios.mallis@nottingham.ac.uk Enrique Sanchez Samsung AI Center, Cambridge, UK e.lozano@samsung.com Matt Bell University of Nottingham matt.bell@nottingham.ac.uk Georgios Tzimiropoulos Queen Mary University of London, UK Samsung AI Center, Cambridge, UK g.tzimiropoulos@qmul.ac.uk
Pseudocode No The paper describes the algorithms and procedures in prose, but does not include any formal pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/malldimi1/Unsupervised Landmarks.
Open Datasets Yes Datasets: We validate our approach on faces, human bodies and cat faces. For faces, we used Celeb A [23], AFLW [18], and the challenging LS3D [4]... For human bodies, we use BBCPose [8] and Human3.6M [14].
Dataset Splits Yes For AFLW we used the official train/test partitions, and for LS3D we followed the same protocol as [4] and used the 300W-LP partition [53] to train our models.
Hardware Specification No The paper does not explicitly state any specific hardware used for running the experiments (e.g., GPU model, CPU type, memory).
Software Dependencies No All models were implemented in Py Torch [28]. ... For K-means, we used the Faiss library [16]. No specific version numbers for PyTorch or Faiss are provided.
Experiment Setup Yes We used RMSprop [13], with learning rate equal to 5 10 4, weight decay 10 5 and batch-size 16 for stage 1 and 64 for stage 2. We set M = 100 and M = 250 clusters for facial and body landmarks, respectively. ... no more than 300,000 iterations are necessary for the algorithm to converge for all datasets.